Paper Reading Library

A searchable collection of paper findings, evidence, limitations, and research insights.

Cross-Model Disagreement as a Label-Free Correctness Signal
read

Matt Gorbett, Suman Jana · 2026

A second model's token-level surprise can identify incorrect answers without labels, especially when the generating model is confidently wrong.

correctness-estimationcross-model-disagreementuncertaintyreasoning

Maintenance workflow

  1. Copy _templates/paper_reading_template.md into learning/paper-reading/.
  2. Name it with a stable identifier, for example 2603.25450-short-title.md.
  3. Record claims together with evidence, experimental conditions, and confidence.
  4. Use a small controlled tag vocabulary so search results remain consistent.